AgriCapsNet: An Attention-Aware Capsule Network for Multicrop Agricultural Landscape Classification Using Hyperspectral Imagery

Fine classification of complex agricultural landscapes is essential for precise agricultural information acquisition. Hyperspectral imagery (HSI) combined with convolutional neural networks (CNNs) offers an effective means for land cover classification. However, conventional CNNs are built on scalar neurons, which limits their ability to effectively filter and capture joint spectral–spatial information in complex scenes. To overcome these limitations, this study proposes the Agricultural Capsule Network (AgriCapsNet), an encoder-decoder capsule network designed for multi-crop agricultural landscape classification with HSI. The encoder integrates a feature-focused former module, comprising lightweight convolution and attention mechanisms, to enhance feature refinement and extraction, together with capsule layers that leverage vector neurons to strengthen hierarchical feature representation. The decoder employs a joint margin and reconstruction loss to improve classification generalization. In addition, a new hyperspectral dataset, Zhang-Jia-Gang, is publicly released to support research on fine-grained agricultural landscape classification. AgriCapsNet was evaluated on three crop-rich HSI datasets and outperformed nine competing models, achieving overall accuracy improvements of 0.90%–24.49%, 1.87%–28.82%, and 1.84%–7.09%, respectively. It also demonstrated strong robustness to noise and adaptability to different pixel neighborhood sizes, confirming its suitability for large-scale agricultural monitoring.